synapse-memory（突触记忆）

Synapse — Self-Learning Memory Engine

Augments OpenClaw's built-in memory system with structured learning, preference tracking, and cross-session intelligence. Does NOT replace memory_search/memory_get — enhances them.

Core Principles

• Use existing OpenClaw memory tools first — memory_search and memory_get are the primary read path
• Synapse adds structure on top — profiles, preferences, patterns, learning logs
• Never fabricate memories — only store what was explicitly stated or directly observed
• Quiet learning — note observations without being asked, but inform the user briefly

Memory Architecture

~/.openclaw/workspace-astra/memory/synapse/
├── profile.json          # Structured operator profile (facts, preferences, patterns)
├── preferences.json      # Tracked preferences with confidence scores
├── patterns.jsonl        # Append-only pattern detection log
├── daily/                # Daily learning cycle outputs
│   └── YYYY-MM-DD.md     # What was learned today
└── associations.json     # Cross-reference map (topic → related memories)

Workflow

On Any Conversation

1. Scan for learnable signals:

- Explicit statements: "I prefer X", "I don't like Y", "Remember that..." - Implicit signals: repeated corrections, consistent tool choices, time-of-day patterns - Decision patterns: chosen option vs rejected alternatives

1. Extract and classify:

- Fact — verifiable statement ("I work at DGA EPM") - Preference — subjective choice ("I prefer TypeScript over Python") - Pattern — behavioral trend ("always asks for cost estimates before builds") - Correction — prior information updated ("actually, it's Teya not Matea in casual context")

1. Store in appropriate file using write or edit tool

Learning Cycles (Triggered by cron or on-demand)

1. Read last 24h of session transcripts (if accessible)
2. Scan MEMORY.md for new entries
3. Check daily memory notes (memory/YYYY-MM-DD.md)
4. Extract new intelligence from steps above
5. Update profile.json with incremental changes
6. Write daily log to daily/YYYY-MM-DD.md
7. Surface summary: "Learned X new facts, Y preferences updated, Z patterns detected"

Recall Flow

1. Always try memory_search first for general queries
2. For profile-specific queries ("what are my preferences?"), read profile.json
3. For pattern queries ("what have I been working on?"), scan daily/ logs
4. Cross-reference associations.json for related topics

Profile Structure (profile.json)

{
  "version": 1,
  "lastUpdated": "ISO-8601",
  "facts": {
    "name": "shadoprizm",
    "location": "Ottawa, Ontario",
    "timezone": "EST"
  },
  "preferences": [
    {
      "category": "communication",
      "item": "direct, no fluff",
      "confidence": 0.95,
      "source": "explicit",
      "firstSeen": "ISO-8601",
      "lastConfirmed": "ISO-8601"
    }
  ],
  "patterns": [
    {
      "description": "Always asks for cost before deploying paid agents",
      "frequency": 12,
      "confidence": 0.9
    }
  ],
  "corrections": [
    {
      "from": "Matea",
      "to": "Teya (casual) / Matea (formal)",
      "date": "ISO-8601"
    }
  ]
}

Constraints

• Do NOT store sensitive information (passwords, API keys, tokens)
• Do NOT fabricate memories — if unsure, note low confidence
• Do NOT replace MEMORY.md — Synapse is a supplement
• Maximum 200 preferences tracked (oldest/lowest-confidence pruned first)
• Daily logs older than 90 days archived to daily/archive/

Integration with OpenClaw Memory

Daily Learning Prompt Template

When running a learning cycle:

Review recent interactions and extract:
1. New facts learned about the operator
2. Preference signals (explicit or implicit)
3. Behavioral patterns observed
4. Corrections to prior knowledge
5. Topics of recurring interest

Update profile.json, write daily/YYYY-MM-DD.md, do NOT fabricate.